 |
 | Home | Contact Us | FAQ | Search & Site Map | Link to Us |
 | Sign In | Join | Other 45 Sites in Network |
.NET Forum / .NET Framework / Performance / February 2004why won't JIT inline this?
I have written a custom replacement for library function Math.Pow(double x, double y) for the specific condition of y being integer and >=0. It is twice as fast as Math.Pow(), but I cannot get the compiler to inline compile it. I've tried many different approaches, but nothing seems to work. I am debugging in native mode, and I do see property get's inlined as I would expect. Any ideas? C# code, IL code, and the optimized disassembly follow: C# code: public static double IntPower(double x, int y) { double r=1; while (y>0) { r*=x; y--; } return r; } ILDASM: .method public hidebysig static float64 IntPower(float64 x, int32 y) cil managed { // Code size 27 (0x1b) .maxstack 2 .locals init (float64 V_0) IL_0000: ldc.r8 1. IL_0009: stloc.0 IL_000a: br.s IL_0015 IL_000c: ldloc.0 IL_000d: ldarg.0 IL_000e: mul IL_000f: stloc.0 IL_0010: ldarg.1 IL_0011: ldc.i4.1 IL_0012: sub IL_0013: starg.s y IL_0015: ldarg.1 IL_0016: ldc.i4.0 IL_0017: bgt.s IL_000c IL_0019: ldloc.0 IL_001a: ret } // end of method EMath::IntPower disassembly: set and call IntPower() function: 02FD04E5 push 40080000h 02FD04EA push 0 02FD04EC mov ecx,2 02FD04F1 call dword ptr ds:[3E6854h] ... IntPower function: 02FD07B0 fld1 02FD07B2 cmp ecx,0 02FD07B5 jle 02FD07C3 02FD07B7 fld qword ptr [esp+4] 02FD07BB fmulp st(1),st 02FD07BD dec ecx 02FD07BE cmp ecx,0 02FD07C1 jg 02FD07B7 02FD07C3 ret 8 > I have written a custom replacement for library function > Math.Pow(double x, double y) for the specific condition of y being [quoted text clipped - 4 lines] > > Any ideas? I believe the JIT compiler doesn't inline anything involving loops.  SignatureJon Skeet - <skeet@pobox.com> http://www.pobox.com/~skeet If replying to the group, please do not mail me too <clipped> > I believe the JIT compiler doesn't inline anything involving loops. That's really disappointing. This loop is so small that it could easily be inlined. I'll manually inline it to see how the performance changes; and also see at what exponent value the calculation time overwhelms the function call overhead. Thanks for the response. >I have written a custom replacement for library function > Math.Pow(double x, double y) for the specific condition of y being [quoted text clipped - 4 lines] > > Any ideas? The JITter will not inline complex flow controls, complex here means anything else then if/then/else (in your case while). Willy. > >I have written a custom replacement for library function > > Math.Pow(double x, double y) for the specific condition of y being [quoted text clipped - 9 lines] > > Willy. SO, you think a 'goto' would be inlined? Something like: IntPower(double x, int y) { r=1; start: if (y>1) { r*=x; y--; goto start; } return r; I think it was said clearly, any flow control other than if/then/else won't be inlined. Goto is not one of if/then/else. Jerry > > >I have written a custom replacement for library function > > > Math.Pow(double x, double y) for the specific condition of y being [quoted text clipped - 21 lines] > } > return r; No, as I said only If/then/else control constructs, nothing that would involve conditional looping. Note that in your sample the call overhead is very small compared to the execution time of the loop, especially when y is large. Willy. >> >I have written a custom replacement for library function >> > Math.Pow(double x, double y) for the specific condition of y being [quoted text clipped - 21 lines] > } > return r; > No, as I said only If/then/else control constructs, nothing that would > involve conditional looping. And what about switch-statements? Signaturecody [Freeware, Games and Humor] www.deutronium.de.vu || www.deutronium.tk Nope. Willy. >> No, as I said only If/then/else control constructs, nothing that would >> involve conditional looping. > > And what about switch-statements? > >> No, as I said only If/then/else control constructs, nothing that would > >> involve conditional looping. > > > > And what about switch-statements? > > Nope. Why not? a switch is in most cases much faster than if..else if.. Signaturecody [Freeware, Games and Humor] www.deutronium.de.vu || www.deutronium.tk But in most cases generates lot more code. JIT won't inline anything over 20 instructions. Jerry > > >> No, as I said only If/then/else control constructs, nothing that would > > >> involve conditional looping. [quoted text clipped - 10 lines] > [Freeware, Games and Humor] > www.deutronium.de.vu || www.deutronium.tk > The JITter will not inline complex flow controls, complex here means > anything else then if/then/else (in your case while). > > Willy. Also avoid using and object or class variables in methods you want to inline. The following code, for instance, can't be inlined: class Optimizer { static int EnergyType = 1; static double Energy(double dx, double dy, double dz, double rp) { return (EnergyType == 1) ? rp*Math.Sqrt(1.0/(dx + dy + dz)) : 1.0; } } CLR JIT can only inline very basic code. We generally inline all our code by copy-and-paste, it is messy, but makes big difference in our case. James visumap.net : Also avoid using and object or class variables in methods you want to : inline. [quoted text clipped - 10 lines] : CLR JIT can only inline very basic code. We generally inline all our : code by copy-and-paste, it is messy, but makes big difference in our case. That's interesting that you can't get this code to inline. I have had similar code inlined (though not with static variables). For example, this type of code inlines in framework 1.0 fine in my app: (note this is a struct, not a class, if that might make some difference?) struct XYZ { private int mVal; static int GetValue() { if mVal < 100 return mVal; else return SomeMoreComplicatedProcessingRequired(); } } Code in previous post was incorrect. GetValue() IS NOT a static function. struct XYZ { private int mVal; public int GetValue() { if mVal < 100 return mVal; else return SomeMoreComplicatedProcessingRequired(); } } Free MagazinesGet these publications absolutely FREE for up to 12 months. There are no hidden fees and no obligation. Simply choose a title, complete the application form and submit it. Read more ...
|
Reply to this Message
|
 Sign In
 Join
 My Latest Posts My Monitored Threads My Blog My Photo Gallery My Profile My Homepage Start New Thread Enable EMail Alerts Rate this Thread
|
©2008 Advenet LLC Privacy Policy - Terms of Use
This website includes both content owned or controlled by Advenet as well as content owned or controlled by third parties.